Research
V-STORM: Structure-Aware Grounded Chain of Thought Dataset for Faithful Video Reasoning
V-STORM is a framework that converts videos into Spatio-Temporal Scene Graphs and derives Grounded Chain of Thought annotations, linking each reasoning step to explicit spatio-temporal evidence. It includes a 260K-sample training set and a 3K-sample human-verified benchmark. Experiments show current multimodal LLMs can localize evidence but fail to compose it into multi-step reasoning, and fine-tuning with V-STORM improves both accuracy and grounding faithfulness.
V-STORMspatio-temporal reasoningmultimodal LLM